Technical skills

• Modeling and statistical learning
• Information and decision theories
• Sensitivity analysis and uncertainty treatment
• Convex, quasi-convex and Bayesian optimization
• Neuro-evolution for neural network architecture optimization
• Multi-agent systems, bandits and game theory
• HPC computing and CPU/GPU optimization

AI by reinforcement learning

Reinforcement learning allows industrialists to develop AIs proposing optimal resource management decisions by adapting to low or non-stationary environments. By hybridizing the description of these environments with historical decision support systems, multiple agent mechanisms and new constraints of delocalized computation, the reinforcement learning algorithms produced in SINCLAIR can be used to test and improve many industrial strategies ranging from maintenance operations to system design.

AI at the service of simulation

Digital simulation, because it allows in silico testing, is an essential lever for large industrial companies. By hybridizing acquired and learned knowledge and producing generative models capable of mimicking real phenomena, such as fluid flows or electromagnetic wave propagation, this line of research feeds directly into the design of augmented engineering tools, such as digital twins.

Trusted AI

In support of the production of models and algorithms, SINCLAIR develops methods and tools (M&O) to understand the nuances of AI behavior. It is indeed essential, for industrial partners, to be able to make their use intelligible, interpretable and explainable. The experience and the culture acquired since a long time on the adoption of M&O produced by the R&D of the industrialists towards engineering facilitate the study and the adaptation of methodologies resulting from the statistics and the analysis of the causality.